electronics

Article
TLI-YOLOv5: A Lightweight Object Detection Framework for
Transmission Line Inspection by Unmanned Aerial Vehicle
Hanqiang Huang 1, * , Guiwen Lan 1,2, *, Jia Wei 1 , Zhan Zhong 1 , Zirui Xu 1 , Dongbo Li 1 and Fengfan Zou 1

1 College of Geomatics and Geoinformation, Guilin University of Technology, Guilin 541006, China
2 Guangxi Key Laboratory of Spatial Information and Geomatics, Guilin University of Technology,
Guilin 541006, China
* Correspondence: [email protected] (H.H.); [email protected] (G.L.)

Abstract: Unmanned aerial vehicles (UAVs) have become an important tool for transmission line
inspection, and the inspection images taken by UAVs often contain complex backgrounds and many
types of targets, which poses many challenges to object detection algorithms. In this paper, we
propose a lightweight object detection framework, TLI-YOLOv5, for transmission line inspection
tasks. Firstly, we incorporate the parameter-free attention module SimAM into the YOLOv5 network.
This integration enhances the network’s feature extraction capabilities, without introducing additional
parameters. Secondly, we introduce the Wise-IoU (WIoU) loss function to evaluate the quality of
anchor boxes and allocate various gradient gains to them, aiming to improve network performance
and generalization capabilities. Furthermore, we employ transfer learning and cosine learning rate
decay to further enhance the model’s performance. The experimental evaluations performed on our
UAV transmission line inspection dataset reveal that, in comparison to the original YOLOv5n, TLI-
YOLOv5 increases precision by 0.40%, recall by 4.01%, F1 score by 1.69%, mean average precision at
50% IoU (mAP50) by 2.91%, and mean average precision from 50% to 95% IoU (mAP50-95) by 0.74%,
while maintaining a recognition speed of 76.1 frames per second and model size of only 4.15 MB,
exhibiting attributes such as small size, high speed, and ease of deployment. With these advantages,
Citation: Huang, H.; Lan, G.; Wei, J.; TLI-YOLOv5 proves more adept at meeting the requirements of modern, large-scale transmission
Zhong, Z.; Xu, Z.; Li, D.; Zou, F. line inspection operations, providing a reliable, efficient solution for such demanding tasks.
TLI-YOLOv5: A Lightweight Object
Detection Framework for Keywords: object detection; SimAM attention module; Wise-IoU (WIoU) loss function; YOLOv5n;
Transmission Line Inspection by
transmission line inspection
Unmanned Aerial Vehicle. Electronics
2023, 12, 3340. https://doi.org/
10.3390/electronics12153340

Academic Editors: Gábor Kertész, 1. Introduction

Sašo Tomažič, Sara Stančin,
With the mounting electricity demands of society, the inspection and maintenance of
Peter Sarcevic and Akos Odry
transmission lines have become increasingly crucial [1,2]. Traditional manual inspection
Received: 18 July 2023 methods are relatively slow, costly, and encompass inherent risks. The utilization of UAVs
Revised: 1 August 2023 for electrical inspections offers many advantages, including safety, efficiency, flexibility,
Accepted: 2 August 2023 cost-effectiveness, and minimal constraints [3–7]. Consequently, UAVs have become widely
Published: 4 August 2023 adopted in power line inspections, becoming the principal tool for the electric utility
sector [8–10].
UAV inspections require the collection of a large volume of image data, and manually
inspecting this vast dataset is time-consuming. Furthermore, the quality of the inspection
Copyright: © 2023 by the authors.
is subject to the subjective judgment and skill level of the personnel. The varying quality
Licensee MDPI, Basel, Switzerland.
of image data could potentially lead to erroneous or missed detections [5]. In recent
This article is an open access article
years, machine vision technology has significantly enhanced inspection efficiency [11].
distributed under the terms and
Jenssen et al. [12] proposed a computer vision-based power line inspection method that
conditions of the Creative Commons
uses UAV optical images as the data source, combined with deep learning algorithms
Attribution (CC BY) license (https://
for data analysis and detection. This approach allows for the automatic detection of
creativecommons.org/licenses/by/
4.0/).
safety risks such as missing pole top pads, incorrectly installed insulators, cracked poles,

Electronics 2023, 12, 3340. https://doi.org/10.3390/electronics12153340 https://www.mdpi.com/journal/electronics

Electronics 2023, 12, 3340 2 of 25

and damage from woodpeckers. Han et al. [13] introduced a computer vision algorithm
that can detect damaged insulator discs and foreign objects lodged between two discs.
Ma et al. [14] proposed a method for the detection of transmission line insulators that
combines UAV imagery with binocular vision perception technologies. This method can
quickly and intelligently detect insulator damage and absences, while also using the global
positioning system (GPS) and UAV flight parameters to calculate the spatial coordinates of
the insulators.
These studies demonstrate that machine vision technology can significantly enhance
the efficiency of inspection operations and is a crucial strategy in the advancement toward
AI-driven power line inspections. However, these models require substantial computational
resources, have large model sizes, and have slow processing speeds, creating significant
challenges for practical deployment. Therefore, the industry is increasingly demanding
lightweight and compact electrical inspection models with high efficiency.
Since 2012, computer vision technology based on deep convolutional neural networks
has rapidly developed. Object detection has become a research hotspot in the field of
computer vision. Object detection algorithms can be categorized into two main classes:
Two-stage and one-stage. Two-stage algorithms involve candidate box generation in the
first stage and accurate target localization in the second stage. Representative algorithms
include R-CNN [15], Faster R-CNN [16], and Mask R-CNN [17]. In contrast, one-stage al-
gorithms predict the target category and location simultaneously, making them suitable for
real-time detection tasks. Notable examples include YOLO [18] and SSD [19]. Thanks to its
speed and accuracy, YOLO quickly gained significant attention in various fields, including
transportation [20–22], agriculture [23–26], epidemic prevention [27,28], geological moni-
toring [29], urban management [30], and medical diagnosis [31,32]. The power industry is
also exploring the application of YOLO algorithms in power line inspection work.
For instance, Chen et al. [33] proposed an electrical component recognition algorithm
framework based on SRCNN and YOLOv3. The SRCNN network is used to perform
super-resolution reconstruction of blurry images to expand the dataset, while YOLOv3
is used to recognize electrical components. Chen et al. [34] used YOLOv3 to propose
a solution for pole detection and counting based on UAV patrol videos, enabling rapid
post-disaster assessment of fallen poles. Tu et al. [35] proposed a model for recognizing
towers and insulators based on an improved YOLOv3 algorithm. The authors removed the
52 × 52 scale feature extraction layer and pruned the three-scale feature extraction layers
down to two to enhance computational speed. Simultaneously, the K-means++ clustering
method was employed to calculate anchor box dimensions, thereby improving detection
accuracy. Zhan [36] created a dedicated dataset for electrical equipment and compared the
detection performance of Faster R-CNN, Mask R-CNN, YOLO, and SSD. Bao et al. [37]
improved the performance of the YOLOv5x model for detecting defects in insulators
and vibration dampers using UAV remote sensing images by incorporating a coordinate
attention (CA) module and replacing the original PANet feature fusion framework with a
bidirectional feature pyramid network (BiFPN).
However, these existing studies mainly employed the early versions of the YOLO
networks, which had inferior comprehensive performance. Even with improvements
to these algorithms, their performance is hardly comparable to the new version of the
YOLO network. Moreover, as the data sets used in these studies usually have lower image
resolution, larger object sizes, and lower background complexity, with only a few types of
objects, they cannot adequately meet the needs of actual complex inspection scenarios.
To resolve the limitations of previous works and better address the demands of real-
world inspection, we have constructed a tailored, high-quality dataset with characteristics
aligned with practical scenarios. On this basis, we propose TLI-YOLOv5, an advanced
lightweight object detection framework specifically designed for transmission line inspec-
tion. The main contributions and innovations of this paper are summarized as follows:
1. We constructed a UAV transmission line inspection dataset. The dataset includes
1231 high-resolution images, with as many as 8 types of labeled targets, and a total
Electronics 2023, 12, 3340 3 of 25

of 39,151 annotated boxes. The images were captured across various provinces and
cities in China, resulting in a dataset with a rich variety of scenarios, a large volume,
and high-quality images. This dataset provides a robust foundation for training
high-quality models.
2. We introduced YOLOv5n to the field of transmission line inspection, a novel appli-
cation that has not been explored before. YOLOv5n, the latest model released in
the YOLOv5v6.0 version, exhibits faster speeds and a smaller size compared to its
predecessors such as YOLOv5s, YOLOv5m, YOLOv5l, and YOLOv5x. These charac-
teristics make it particularly well-suited for large-scale, real-time transmission line
inspection tasks.
3. We constructed the TLI-YOLOv5, a lightweight object detection framework for trans-
mission line inspection, built upon the foundation of YOLOv5. Firstly, we incorpo-
rated the parameter-free attention module, SimAM, into the backbone of the YOLOv5n
network, which enables a bolstered feature extraction ability without increasing the
network parameters. Secondly, the loss function was improved to Wise-IoU (WIoU),
enhancing the model’s accuracy, robustness, and generalization ability. Furthermore,
we employed transfer learning techniques to expedite the convergence rate during
training and augment the learning performance of the model. We also adopted a
cosine learning rate decay strategy to ensure a more stable training process, hasten the
convergence speed, and thus improve the better generalization ability of the model.
4. We validated the proposed TLI-YOLOv5 on our transmission line inspection dataset.
The experimental evaluations revealed that, in comparison to the original YOLOv5n
model, the proposed TLI-YOLOv5 model exhibited measurable improvements. Specif-
ically, precision improved by 0.40%, recall increased by 4.01%, the F1 score rose by
1.69%, the mean average precision at 50% IoU (mAP50) increased by 2.91%, and the
mean average precision from 50% to 95% IoU (mAP50-95) also increased by 0.74%.
Moreover, the model maintained a recognition speed of 76.1 frames per second (FPS)
and a compact size of only 4.15 MB.
The remaining sections of this paper are structured as follows. In Section 2, we
provide a detailed exposition on the construction of our UAV transmission line inspection
dataset, along with a comprehensive elaboration on the architecture and principles of TLI-
YOLOv5. This includes the principles and integration of YOLOv5n, the SimAM attention
module, and the WIoU loss function. We also describe the training methods involving
transfer learning and cosine learning rate decay. In Section 3, we present a comprehensive
description of the rigorous experimental procedures undertaken. Section 4 is dedicated to a
thorough discussion of the strengths and limitations of the proposed framework. Finally, in
Section 5, we provide a concise summary that outlines the main findings and contributions
of our study.

2. Materials and Methods

2.1. Dataset Construction
2.1.1. Transmission Line Inspection Dataset
The original imagery for the dataset used in this study was provided by a large-scale
power grid company and comprises 13,744 real electrical inspection images. Adhering to
the principles of maximizing scene diversity and image clarity, we selected 1231 images for
annotation, and manually delineated 39,151 bounding box labels.
The image collection covers a wide range of regions across multiple provinces and
cities in China and is characterized by high image quality and complex backgrounds, pro-
viding robust support for training high-quality models. The dataset encompasses a myriad
of settings, including but not limited to urban landscapes, rural environs, mountainous
terrains, and forested areas, as shown in Figure 1.
Electronics 2023, 12, x FOR PEER REVIEW 4 of 26

providing robust support for training high-quality models. The dataset encompasses a
Electronics 2023, 12, 3340 4 of 25
myriad of settings, including but not limited to urban landscapes, rural environs, moun-
tainous terrains, and forested areas, as shown in Figure 1.

Figure 1.
Figure 1. Showcase of the
Showcase of the diverse
diverse scenarios
scenarios covered
covered in
in the
the dataset.
dataset.

The captured images were taken at a diverse array of distances, angles, and weather
conditions, as
conditions, asdepicted
depictedininFigure
Figure2. 2. This
This diverse
diverse collection
collection of samples
of samples is specifically
is specifically cu-
curated
rated
to to ensure
ensure a broad
a broad spectrum
spectrum of scenarios
of scenarios and and enhance
enhance the the generalization
generalization capabilities
capabilities of
of the
the model.
model.
In alignment with the practical requirements of inspection work, we annotated eight
types of targets, including towers, insulators, yoke plates, corona rings, line clamps, vi-
bration dampers, tower signs, and bird nests, as shown in Figure 3. To make the model’s
output display more neat and easily interpretable, we abbreviated the names of some of the
targets. The final annotation labels used for the eight types of targets are as follows: tower,
insulator, plate, ring, clamp, damper, sign, and nest. Finally, we partitioned the dataset into
three subsets: the training set, validation set, and test set, with a ratio of 8:1:1.

2.1.2. Dataset Annotation

For annotation, we employed the annotation tool labelImg, with the annotation process
depicted in Figure 4. Due to the diversity of the categories and the wide viewing angles
captured by the drones, a single image could contain dozens of bounding boxes. Across the
1231 images, a total of 39,151 bounding boxes were annotated, averaging approximately
31.8 bounding boxes per image.
Electronics 2023, 12, 3340 5 of 25
Electronics 2023, 12, x FOR PEER REVIEW 5 of 26

Figure 2. Display of the dataset captured at various distances, angles, and weather conditions: (a)
2. Display
FigureImagery of at
captured the dataset
different captured
distances; at various
(b) imagery distances,
captured angles,
from different and
angles; (c) weather
imagery conditions:
Electronics 2023, 12, x FOR PEER REVIEW 6 of 26
(a) Imagery captured
captured at different
under different distances;
weather/lighting (b) imagery captured from different angles;
conditions. (c) imagery
captured under different weather/lighting conditions.
In alignment with the practical requirements of inspection work, we annotated eight
types of targets, including towers, insulators, yoke plates, corona rings, line clamps, vi-
bration dampers, tower signs, and bird nests, as shown in Figure 3. To make the model’s
output display more neat and easily interpretable, we abbreviated the names of some of
the targets. The final annotation labels used for the eight types of targets are as follows:
tower, insulator, plate, ring, clamp, damper, sign, and nest. Finally, we partitioned the
dataset into three subsets: the training set, validation set, and test set, with a ratio of 8:1:1.

Figure 3. Examples of the 8 types of targets: (a) tower; (b) insulator; (c) yoke plate; (d) corona ring;
Figure 3. Examples of the 8 types of targets: (a) tower; (b) insulator; (c) yoke plate; (d) corona ring;
(e) line clamp; (f) vibration damper; (g) tower sign; (h) bird nest.
(e) line clamp; (f) vibration damper; (g) tower sign; (h) bird nest.
2.1.2. Dataset Annotation
For annotation, we employed the annotation tool labelImg, with the annotation pro-
cess depicted in Figure 4. Due to the diversity of the categories and the wide viewing
angles captured by the drones, a single image could contain dozens of bounding boxes.
Across the 1231 images, a total of 39,151 bounding boxes were annotated, averaging ap-
proximately 31.8 bounding boxes per image.
Electronics 2023, 12, x FOR PEER REVIEW 7
Electronics 2023, 12, 3340 6 of 25

Figure4.4.Demonstration
Figure Demonstrationof the
ofannotation method.
the annotation method.

A statistical overview of the dataset is presented in Figure 5. Figure 5a depicts the

A statistical overview of the dataset is presented in Figure 5. Figure 5a depicts
distribution of bounding box annotations across the eight object categories in the dataset,
distribution of bounding
providing an overview box annotations
of the classes acrossquantities.
and their respective the eightFigure
object5bcategories in the dat
illustrates the
providing
size an overview
of each annotated of the
bounding boxclasses and their
in the dataset, respective
revealing that a quantities. Figure 5b illustr
substantial proportion
the size of each annotated bounding box in the dataset, revealing that a substantial
of the objects in the dataset are of small size. This presents a considerable challenge for
object
portiondetection
of thealgorithms.
objects inFigure 5c provides
the dataset are aofplot of the
small normalized
size. center coordinates
This presents a considerable c
of each bounding box, remapped to a range of 0–1, displaying the spatial distribution of the
lenge for object detection algorithms. Figure 5c provides a plot of the normalized ce
objects. Figure 5d presents the width-to-height ratio of each bounding box, also normalized
coordinates
to the range ofof each
0–1. Thebounding box, points
notably darker remapped
in the to a range
lower of 0–1,
left corner displaying
further emphasizethe spatial
tribution
the of the objects.
high proportion Figure
of smaller 5d in
targets presents the width-to-height ratio of each bounding
the dataset.
also normalized to the range of 0–1. The notably darker points in the lower left co
further emphasize the high proportion of smaller targets in the dataset.
Electronics 2023, 12, x FOR PEER REVIEW 8 of 26
Electronics 2023, 12, 3340 7 of 25

Figure 5. Statistical overview of the dataset: (a) Number of instances per class; (b) sizes of bounding
boxes;
Figure (c) normalized
5. Statistical bounding
overview boxes
of the centers;
dataset: (a)(d) normalized
Number width and
of instances perheight
class;of(b)
bounding
sizes ofboxes.
bounding
boxes;
2.2.(c)
A normalized bounding
Brief Introduction boxes centers; (d) normalized width and height of bounding boxes.
to YOLOv5
2.2.1. The Network Architecture of YOLOv5
2.2. A Brief Introduction to YOLOv5
In June 2020, Glenn Jocher unveiled YOLOv5 [38], which has since undergone con-
2.2.1. The Network
tinuous Architecture
optimization of YOLOv5
and refinement and is still actively being updated. The network
architecture of YOLOv5
In June 2020, is illustrated
Glenn Jocher in Figure
unveiled 6.
YOLOv5 [38], which has since undergone con-
tinuous optimization and refinement and is still actively being updated. The network ar-
chitecture of YOLOv5 is illustrated in Figure 6.
Electronics 2023,
Electronics 12, 12,
2023, 3340x FOR PEER REVIEW 9 of 826of 25

Figure6.6.Network
Figure Networkarchitecture
architecture of
of YOLOv5.
YOLOv5.

TheYOLOv5
The YOLOv5modelmodel is is structured
structuredinto intothree
threeprimary
primarysections:
sections:thethe
backbone,
backbone, thethe
neck,
neck,
and the head.
and the head.
Thebackbone
The backboneformsformsthe thecore
coreofofthe
thenetwork
network and
and is is responsible
responsible for for the
the initial
initial stages
stages of
feature extraction. It is built upon the new CSP-Darknet53 structure, an evolved version
of feature extraction. It is built upon the new CSP-Darknet53 structure, an evolved version of
of Darknet
the the Darknet architecture
architecture fromfrom earliermodels.
earlier models. The
The backbone
backboneincludesincludesCBS,CBS, C3,
C3,andandSPPFSPPF
modules. The convolution, batch normalization, and SiLU (CBS)
modules. The convolution, batch normalization, and SiLU (CBS) module is a fundamental module is a fundamental
building block in the YOLOv5 architecture. It begins with a convolution operation to ex-
building block in the YOLOv5 architecture. It begins with a convolution operation to
tract specific features from the input, followed by batch normalization to standardize
extract specific features from the input, followed by batch normalization to standardize
these features and increase the stability of the network. Finally, the SiLU (sigmoid linear
these features and increase the stability of the network. Finally, the SiLU (sigmoid linear
unit) activation function introduces non-linearity, allowing the model to learn and repre-
unit) activation function introduces non-linearity, allowing the model to learn and represent
sent more complex patterns. The C3 module, standing for “CSP Bottleneck with 3 convo-
more complex patterns. The C3 module, standing for “CSP Bottleneck with 3 convolutions”,
lutions”, further enhances the feature extraction process by splitting the input feature map
further enhances the feature extraction process by splitting the input feature map into two
into two parts, applying a sequence of convolution layers, and then merging the results.
parts, applying a sequence of convolution layers, and then merging the results. The SPPF
The SPPF module, a variant of the spatial pyramid pooling (SPP) module, extracts features
module, a variant
at different scales,ofcrucial
the spatial pyramidobjects
for handling pooling
of (SPP)
various module,
sizes inextracts
the input features
images. at different
scales,The
crucial for handling objects of various sizes in the input images.
neck of the YOLOv5 model connects the backbone and the head. It utilizes the
newThe neck ofstructure
CSP-PAN the YOLOv5 model
to further connects
process the the backbone
features and by
extracted thethe
head. It utilizes
backbone. Thethe
new CSP-PAN structure to further process the features extracted by
neck is responsible for consolidating the high-level semantic feature maps from the back- the backbone. The neck
isbone
responsible for consolidating the high-level semantic feature maps
and combining them in a way that preserves both the high-resolution details and the from the backbone
and combining
broader context.them in a way that preserves both the high-resolution details and the
broader context.
The head of the YOLOv5 model is responsible for generating the final output, i.e., the
Thedetection
object head of predictions.
the YOLOv5Itmodel uses the is responsible
YOLOv3 head forstructure,
generating the final
which output,
consists i.e., the
of multi-
object detection predictions.
ple convolutional layers andItoutput uses the YOLOv3
layers. head structure,
The output which
layers predict theconsists of multiple
class probabili-
convolutional
ties and boundinglayersboxand output layers.
coordinates for eachThe output
grid cell inlayers predict
the feature map.theThese
classpredictions
probabilities
and
arebounding box coordinates
then processed to produce the for final
eachobject
grid cell in the results.
detection feature map. These predictions are
then processed to produce the final object detection results.
YOLOv5 introduces several modifications compared to its predecessors, such as the
replacement of the Focus structure with a 6 × 6 Conv2d structure for enhanced efficiency
and the substitution of the SPP structure with SPPF to accelerate processing speed.
Electronics 2023, 12, 3340 9 of 25

To enhance the model’s generalization capabilities and minimize overfitting, YOLOv5

employs a diverse range of data augmentation techniques. These include mosaic augmen-
tation, copy-paste augmentation, random affine transformations, MixUp augmentation,
Albumentations, HSV augmentation, and random horizontal flip. Each of these meth-
ods plays a crucial role in improving the model’s adaptability to various object scales,
translations, and image conditions, thereby boosting its overall performance.
YOLOv5 also utilizes a series of sophisticated training strategies to optimize the
model’s performance. These include multiscale training, AutoAnchor, exponential moving
average (EMA), mixed precision training, and hyper-parameter evolution. Collectively,
these strategies optimize the learning process, stabilize training, reduce generalization
error, and ensure optimal performance.

2.2.2. Overview of YOLOv5n

The YOLOv5 model is available in different variants: YOLOv5n, YOLOv5s, YOLOv5m,
YOLOv5l, and YOLOv5x. These variants are designed to offer a range of trade-offs between
computational efficiency and detection performance, allowing users to choose the variant
that best suits their specific needs.
YOLOv5n, introduced as a lightweight model in version 6.0 of YOLOv5 [39], is the
smallest and fastest model among the YOLOv5 variants. It is designed for applications
where computational resources are limited, making it ideal for deployment on devices
with low computational power. The compactness of YOLOv5n is achieved by setting lower
values for two key hyper-parameters: depth multiple and width multiple.
The depth multiple is a hyper-parameter that controls the depth of the network, which
corresponds to the number of layers in the model. A lower-depth multiple results in fewer
layers, thereby reducing the computational complexity of the model and making it faster to
run. However, this also means that the model may extract fewer features from the input
data, which could potentially impact the detection performance.
The width multiple is another hyper-parameter that controls the width of the net-
work, corresponding to the number of channels in each layer. A lower width multiple
results in fewer channels, further reducing the computational complexity of the model.
However, similar to the depth multiple, a lower width multiple may limit the amount of
information that the model can capture in each layer, which could also potentially impact
the detection performance.
By adjusting these hyper-parameters, it is possible to modify the network’s depth and
width without altering the underlying architecture. This results in five variants—YOLOv5n,
YOLOv5s, YOLOv5m, YOLOv5l, and YOLOv5x—each with varying levels of depth and
width. A comparative overview of the depth and width hyper-parameters across these five
models is provided in Table 1.

Table 1. The depth and width parameters of the five YOLOv5 variants.

YOLOv5n YOLOv5s YOLOv5m YOLOv5l YOLOv5x

Depth multiple 0.33 0.33 0.67 1.00 1.33
Width multiple 0.25 0.50 0.75 1.00 1.25

YOLOv5n is particularly advantageous for transmission line inspections due to its

computational efficiency and speed, enabling real-time detection crucial for timely issue
resolution. Its lower computational demands make it ideal for deployment on resource-
constrained devices such as drones, commonly used in these inspections. Furthermore, its
efficiency facilitates long-term, high-frequency monitoring, leading to significant energy
and resource savings. In emergency scenarios, YOLOv5n’s ability to rapidly process data is
critical for prompt issue identification and response. While other YOLOv5 variants may
offer higher detection performance, their increased computational requirements make them
less suitable for power line inspections, particularly on devices with limited computational
Electronics 2023, 12, x FOR PEER REVIEW 11 of 26

Electronics 2023, 12, 3340 is critical for prompt issue identification and response. While other YOLOv5 variants 10may
of 25
offer higher detection performance, their increased computational requirements make
them less suitable for power line inspections, particularly on devices with limited compu-
tational power. Therefore, YOLOv5n’s balance of efficiency and performance makes it the
power. Therefore, YOLOv5n’s balance of efficiency and performance makes it the preferred
preferred choice for these tasks.
choice for these tasks.
2.3. The
2.3. The Proposed
Proposed TLI-YOLOv5
TLI-YOLOv5
UAV transmission
UAV transmission lineline inspection
inspection images
images are
are characterized
characterized byby high
high image
image resolution,
resolution,
high background
high complexity, small
background complexity, small detection
detection targets,
targets, and
and aa variety
variety of
of target types, posing
target types, posing
significant challenges for real-time multi-class detection. To address these challenges,
significant challenges for real-time multi-class detection. To address these challenges, we we
propose TLI-YOLOv5, a lightweight object detection framework specifically designed for
propose TLI-YOLOv5, a lightweight object detection framework specifically designed for
transmission line
transmission line inspection.
inspection. The Theoverall
overallframework
framework ofof
our proposed
our proposed TLI-YOLOv5
TLI-YOLOv5 is il-
is
lustrated ininFigure
illustrated Figure7.7.ItItisismainly
mainlycomprised
comprisedofofthethefollowing
followingcomponents.
components.

Figure 7.
Figure 7. TLI-YOLOv5
TLI-YOLOv5framework:
framework:(a)(a)
Improved YOLOv5
Improved YOLOv5 network: the last
network: the C3
lastmodule of theof
C3 module back-
the
bone is modified to C3SimAM; (b) enhanced loss function: CIoU is refined to WIoU; (c) training
backbone is modified to C3SimAM; (b) enhanced loss function: CIoU is refined to WIoU; (c) training
strategy: implements both transfer learning and cosine learning rate decay strategies.
strategy: implements both transfer learning and cosine learning rate decay strategies.

Firstly, we introduce the SimAM attention module [40] into the YOLOv5n network to
enhance its feature extraction capabilities. Specifically, we incorporate the SimAM attention
Electronics 2023, 12, 3340 11 of 25

module into the C3 module to build the C3SimAM module, which replaces the last C3
module in the original YOLOv5 backbone. Additionally, we improve the model’s perfor-
mance by enhancing the loss function using WIoU [41]. To optimize the training process,
we employ two effective strategies: transfer learning [42–44] and cosine learning rate de-
cay [45]. These strategies contribute to overall improved detection performance and enable
the model to effectively handle the intricacies involved in transmission line inspection.

2.3.1. SimAM Attention Module

In the field of computer vision, a neuron refers to a node in a neural network that
receives input from the previous layer and passes it on to the next. In a neural network,
each neuron corresponds to a specific feature, such as edges, textures, colors, etc. In
visual neuroscience, the most informative neurons are often those that exhibit discharge
patterns that are distinct from surrounding neurons. Inspired by this, Yang et al. proposed
the SimAM attention module in 2021 [40]. The authors considered neurons that exhibit
significant spatial suppression effects to be more important and assigned them higher
priority. To identify these neurons, the authors proposed a simple implementation method
that measures the linear separability between a target neuron and other neurons. To
this end, they defined an energy function to find neurons that exhibit significant spatial
suppression effects and assigned them higher priority. Unlike existing channel-domain
and spatial-domain attention modules, this module can infer three-dimensional attention
weights in the feature map without increasing the parameters in the original network.
The transmission line inspection images captured by UAVs contain complex back-
grounds and multiple types of objects. To accurately detect objects of interest, the model
needs to learn discriminative features that can distinguish between objects and background
areas. However, the original YOLOv5 backbone may not be able to sufficiently extract
such discriminative features. To address this issue, we incorporated the SimAM attention
module into the YOLOv5 backbone.
For transmission line inspection images, each neuron in the feature maps of the
YOLOv5 backbone may respond to different basic visual patterns. By measuring the linear
separability between a target neuron and its surrounding neurons, SimAM can identify
the most distinctive neurons within each channel, which likely correspond to informative
visual patterns critical for distinguishing objects in complex environments. The 3D attention
weights generated by SimAM then allow for selectively highlighting these informative
neurons while suppressing irrelevant ones. This enhances YOLOv5’s capability to extract
discriminative features from transmission line inspection images.
We conducted a series of experiments to determine the optimal way to integrate the
SimAM attention module into the YOLOv5 network. Initially, we attempted to add the
SimAM attention module as an additional layer at various positions within the YOLOv5
backbone, but these modifications did not result in significant performance improve-
ments. So, we tried to combine the SimAM attention module with the C3 module, the i.e.,
C3SimAM module. The structure of the C3SimAM module is depicted in Figure 8. We
then replaced some C3 modules within the backbone with the C3SimAM module. Optimal
performance was achieved when only the last C3 module in the backbone was replaced.
Our experiments validate the effectiveness of this improvement, significantly enhanc-
ing the network’s feature extraction capability. Detailed experimental results and analyses
are presented in Section 3.3.4.
The SimAM attention module mainly consists of two steps:
1. Calculate the minimum energy e*t with Equation (1).

4 σ̂2 + λ


e*t = (1)
(t − µ̂)2 + 2σ̂2 + 2λ
2
where µ̂ = M1
∑iM 2 1 M
=1 xi , σ̂ = M ∑i =1 ( xi − µ̂ ) , t and x are the target neuron and other
neurons within a single channel of the input features, respectively. i is the index number
Electronics 2023, 12, 3340 12 of 25

in the spatial dimension, M is the number of neurons in the channel, and λ is a weight
constant.
From this, it can be seen that the smaller e*t , the larger the difference between t and x,
indicating that the importance of the target neuron in the channel is higher, and its weight
1/e*t is also larger.
2. Enhance the input features are enhanced with Equation (2).

∼
 
1
X = sigmoid X (2)
E
∼
Electronics 2023, 12, x FOR PEER REVIEWIn this equation, X denotes the input features, X denotes the output features, 13
E denotes
of 26
the combination of e*t across all channels and spatial dimensions.

Figure8.8.Schematic
Figure Schematic diagram
diagram of
of the
the C3SimAM
C3SimAMmodule
modulestructure.
structure.

The SimAM
2.3.2. Selection attention
of Loss module mainly consists of two steps:
function
1. TheCalculate
default thebounding
minimum box energy 𝑒 ∗ with Equation
regression (1). in YOLOv5 is complete IoU
loss function
(CIoU) [46]. The CIoU loss function takes into account the differences between the ground
4(𝜎 + 𝜆)
truth and predicted bounding boxes 𝑒 ∗ in
= terms of overlap area, center distance, and aspect (1) ra-
(𝑡 − 𝜇̂ ) + 2𝜎 + 2𝜆
tio. Compared to the traditional IoU loss function, the CIoU loss function exhibits superior
where 𝜇̂ = and
performance ∑ stability,
𝑥 , 𝜎 =thus ∑ being(𝑥 −widely
𝜇̂ ) , t and
used in various
x are object
the target detection
neuron algorithms.
and other neu-
However,
rons withinthere still exists
a single channelroom for input
of the optimization
features,inrespectively.
the CIoU loss i isfunction.
the index number in
the In 2023,dimension,
spatial Tong et al.M [41] proposed
is the numbera of loss function
neurons withchannel,
in the a dynamicand focusing mechanism
λ is a weight con-
called
stant.Wise-IoU (WIoU). The WIoU loss function evaluates the quality of anchor boxes by
calculating
From their
this, itoutlier
can bedegree
seen thatandthe smaller 𝑒 ∗ ,allocates
accordingly the largerdifferent gradient
the difference gains.t Specifi-
between and
cally, for anchor
x, indicating thatboxes with a small
the importance ofoutlier
the target degree, corresponding
neuron in the channel to high-quality
is higher, andanchor
its
boxes, a ∗
smaller gradient gain is assigned to prevent overfitting and enhance the model’s
weight 1/𝑒 is also larger.
generalization
2. Enhance the ability.
inputFor anchorare
features boxes with awith
enhanced large outlier (2).
Equation degree, corresponding to low-
quality anchor boxes, a smaller gradient gain is assigned to mitigate the harmful gradients
induced by low-quality examples, thereby 1 the model’s accuracy and robustness.
improving
𝑋 = 𝑠𝑖𝑔𝑚𝑜𝑖𝑑 ⊙𝑋 (2)
𝐸
For anchor boxes with a moderate outlier degree, corresponding to ordinary-quality anchor
boxes,Ina this
larger gradient
equation, 𝑋 gain
denotesis assigned
the inputto 𝑋 denotes
concentrate
features, the bounding
the output box 𝐸 de-task
regression
features,
on these
notes theanchor boxes. of
combination 𝑒 ∗ across allthe
Furthermore, classification
channels and spatialstandard for anchor box quality is
dimensions.
dynamic, allowing the WIoU loss function to continually determine the most appropriate
2.3.2. Selection
gradient of Loss function
gain allocation strategy under the current circumstances.
The default bounding box regression loss function in YOLOv5 is complete IoU
(CIoU) [46]. The CIoU loss function takes into account the differences between the ground
truth and predicted bounding boxes in terms of overlap area, center distance, and aspect
ratio. Compared to the traditional IoU loss function, the CIoU loss function exhibits supe-
rior performance and stability, thus being widely used in various object detection algo-
rithms. However, there still exists room for optimization in the CIoU loss function.
In 2023, Tong et al. [41] proposed a loss function with a dynamic focusing mechanism
called Wise-IoU (WIoU). The WIoU loss function evaluates the quality of anchor boxes by
Electronics 2023, 12, 3340 13 of 25

The WIoU loss can be calculated by Equation (3).

!
ρ2 b, b gt
LW IoU = r · exp ∗ · (1 − IoU ) (3)
( c2 )

(1− IoU )∗
where r = δα β−δ , β = 1− IoU ∈ [0, +∞), IoU stands for the Intersection over Union
β

between the ground truth and predicted bounding boxes, r is the non-monotonic focusing
coefficient, β is the outlier degree, and α and δ are hyper-parameters, the superscript * de-
notes the operation of detaching variables
from the computation graph to prevent gradients
that could impede convergence, ρ b, b gt represents the Euclidean distance between the
centers of the ground truth and predicted bounding boxes, c represents the diagonal dis-
tance of the smallest rectangular area that can contain both the ground truth and predicted
bounding boxes.
We performed a series of experiments to determine the optimal selection of α and
δ hyper-parameters for the WIoU loss function. The experimental results indicated that
the model performed optimally when α and δ were set to 1.9 and 3, respectively. Detailed
experimentations related to this are presented in Section 3.3.1.
Furthermore, we conducted a series of experiments that demonstrated the positive
impact of the WIoU loss function within our proposed framework. The experimental
results show that the model trained using the WIoU loss function exhibits better overall
performance than the model trained using the CIoU loss function. Relevant experimental
results and analyses can be found in Section 3.3.4.

2.3.3. Transfer Learning

Transfer learning is a method that enhances the detection capability and training speed
of a model in a target domain by leveraging features learned from a pre-trained network
in a similar domain. Transfer learning can reduce the dependency on the target domain
dataset, accelerate training speed, and prevent overfitting, among other benefits.
In this study, we utilize the COCO dataset as the source domain and input it into
YOLOv5n to obtain pre-trained weights. These pre-trained weights, along with the trans-
mission line inspection dataset, are then input into the YOLOv5n model integrated with
the Simam attention module and WIoU loss function to obtain the final weights.
The COCO dataset comprises over 330,000 images with more than 2 million annotated
boxes, covering 80 common object categories in daily life. The images in the COCO dataset
are diverse in terms of angles and backgrounds. Using the weight file trained on this
dataset as the source domain for transfer learning can positively contribute to the training
of the transmission line inspection model.
The effectiveness of this transfer learning approach is verified in our experiments,
which can be found in Section 3.3.2.

2.3.4. Learning Rate Adjustment

Different learning rate adjustment strategies have been employed in various versions
of YOLOv5. Up until version v6.0, the default adjustment method incorporated a combi-
nation of warmup and cosine learning rate decay (cos-lr). Starting with version v6.1 and
in subsequent iterations, the default strategy was changed to warmup and linear learning
rate decay (linear-lr).
Each of these strategies has distinct advantages and disadvantages that must be
considered within the context of the specific task. The cosine learning rate decay strategy
applies a cosine function to decay the learning rate, leading to smoother convergence and
better fine-tuning of the model parameters towards the end of the training. Conversely,
the linear learning rate decay strategy applies a constant rate of decay, providing more
predictable training dynamics and potentially benefiting cases where the initial learning
rate has been finely tuned.
Electronics 2023, 12, x FOR PEER REVIEW 15 of 26

Electronics 2023, 12, 3340

the linear learning rate decay strategy applies a constant rate of decay, providing 14 more
of 25
predictable training dynamics and potentially benefiting cases where the initial learning
rate has been finely tuned.
In our experiments, we have critically assessed both strategies within our detection
In our experiments, we have critically assessed both strategies within our detection
task. Our results, detailed in Section 3.3.3, led us to opt for the cosine learning rate decay
task. Our results, detailed in Section 3.3.3, led us to opt for the cosine learning rate decay
strategy for our final model due to its superior performance in our task.
strategy for our final model due to its superior performance in our task.
2.4.
2.4.Research
ResearchMethodology
MethodologyOverview
Overview
Figure
Figure99provides
provides aa comprehensive
comprehensive and and concise
concise representation
representationofofthe
theresearch
researchmethod-
meth-
odology
ology and andworkflow
workflowemployed
employedininour ourstudy.
study. Specifically,
Specifically, the
the process
process began
began with
with the
the
annotation of eight types of targets on real transmission line inspection images,
annotation of eight types of targets on real transmission line inspection images, resulting in resulting
in
thethe construction
construction of aof a transmission
transmission line inspection
line inspection dataset.dataset. Subsequently,
Subsequently, the TLI-
the TLI-YOLOv5
YOLOv5
framework was developed by integrating the parameter-free attention module,module,
framework was developed by integrating the parameter-free attention SimAM,
SimAM,
the WIoU theloss
WIoU loss function,
function, and advanced
and advanced training strategies
training strategies into the YOLOv5n
into the YOLOv5n network. net-
This
work. This framework
framework resulted inresulted in the realization
the realization of a lightweight
of a lightweight objectmethod
object detection detection method
tailored for
tailored for transmission
transmission line inspection.
line inspection.

Flowchartofofour
Figure9.9.Flowchart
Figure ourstudy
studymethod.
method.

3. Results
3. Results
3.1. Experiment Configuration
3.1. Experiment Configuration
To validate the effectiveness of the TLI-YOLOv5 framework, we carried out a series of
To validate
experiments. thedetailed
The effectiveness of the TLI-YOLOv5
configuration framework,
of the experimental we carriedisout
environment a series
outlined in
of experiments.
Table 2. The detailed configuration of the experimental environment is outlined in
Table 2.
Table 2. System configuration and software environment.
Table 2. System configuration and software environment.
Component Specification
ComponentCPU Specification
Intel(R) Xeon(R) Gold 6240 CPU @ 2.60 GHz
CPU RAM Intel(R) Xeon(R) Gold 6240 128 CPU
GB @ 2.60 GHz
RAM GPU NVIDIA128Quadro
GB RTX5000 16 GB
Operating system Windows10 pro
GPU NVIDIA Quadro RTX5000 16 GB
CUDA version 10.2
Operating system
cuDNN version Windows10 pro
7.6.5
CUDA versionframework
Deep learning 10.2
PyTorch 1.8.2
cuDNN Python version
version 7.6.5 3.8.8
Deep learning framework PyTorch 1.8.2
When Python version
inputting images into the YOLOv5 network, 3.8.8 the images are adjusted to a
predetermined size, with the default size being 640 pixels. Considering that the images
from the transmission line inspection dataset used in this study all have a resolution
exceeding 4000 × 3000 pixels, substantial image compression occurs when these images are
input into the network. This compression can result in significant information loss, leading
Electronics 2023, 12, 3340 15 of 25

to suboptimal model performance. To balance model performance and computational

resources, we have chosen an input size of 1280 pixels, a batch size of 4, and 300 epochs
for training.

3.2. Evaluation Metrics

In this study, the assessment of the model’s performance was conducted using com-
monly utilized metrics in object detection tasks, which include precision (P), recall (R), F1
score (F1), mean average precision at 50% IoU (mAP50), and mean average precision from
50% to 95% IoU (mAP50-95). Furthermore, we adopt frames per second (FPS) as the metric
for evaluating the speed of the model.
Precision, also known as positive predictive value, measures the proportion of true pos-
itives (TP) out of the sum of true positives and false positives (FP), which are instances incor-
rectly identified as positive. The computation formula for precision is shown in Equation (4).
Precision serves to evaluate how many of the detected objects are correct detections.

TP
P= (4)
( TP + FP)

Recall, sometimes referred to as sensitivity or true positive rate, quantifies the propor-
tion of true positives (TP) out of the sum of true positives and false negatives (FN), which
are positive instances incorrectly identified as negative. The computation formula for recall
is presented in Equation (5). Recall is utilized to evaluate how well the model can identify
all positive instances.
TP
R= (5)
( TP + FN )
The F1 score is the harmonic mean of precision and recall. It aims to balance precision
and recall by providing a single metric that encapsulates the model’s performance in terms
of both precision and recall.
The mean average precision at 50% IoU (mAP50) is computed by averaging the
precision-recall curve’s area under the curve across all classes at an IoU threshold of 50%.
The mean average precision from 50% to 95% IoU (mAP50-95) is calculated by aver-
aging the mAP values over IoU thresholds from 0.50 to 0.95 (inclusive) with a step size
of 0.05.

3.3. Experimental Results and Analysis

3.3.1. Experiment on WIoU Hyper-Parameter Selection
In paper [41], the authors enumerated four sets of values for the hyper-parameters α
and δ, which are α = 2.5, δ = 2; α = 1.9, δ = 3; α = 1.6, δ = 4; and α = 1.4, δ = 5. The outlier
degree β and gradient gain r curves corresponding to these four sets of values 17
Electronics 2023, 12, x FOR PEER REVIEW areofdepicted
26
in Figure 10.

Figure 10. Outlier degree 𝛽 and gradient gain 𝑟 curves for different 𝛼 and 𝛿 hyper-parameters.
Figure 10. Outlier degree β and gradient gain r curves for different α and δ hyper-parameters.
In the context of the TLI-YOLOv5 framework, we have tested the model performance
for these four sets of hyper-parameters, as shown in Table 3. From these experiments, we
have determined the optimal values of 𝛼 and 𝛿 to be 1.9 and 3, respectively.

Table 3. Model performance for different 𝛼 and 𝛿 hyper-parameters.

Electronics 2023, 12, 3340 16 of 25

In the context of the TLI-YOLOv5 framework, we have tested the model performance
for these four sets of hyper-parameters, as shown in Table 3. From these experiments, we
have determined the optimal values of α and δ to be 1.9 and 3, respectively.

Table 3. Model performance for different α and δ hyper-parameters.

Values of α and δ P R F1 mAP50 mAP50-95

α = 2.5 δ = 2 0.736 0.503 0.59 0.545 0.271
α = 1.9 δ = 3 0.752 0.519 0.60 0.566 0.271
α = 1.6 δ = 4 0.770 0.497 0.60 0.554 0.265
α = 1.4 δ = 5 0.713 0.497 0.58 0.541 0.266

3.3.2. Experiment on the Effectiveness of Transfer Learning

To validate the effectiveness of the transfer learning strategy, we conducted two sets
of experiments within the TLI-YOLOv5 framework. As shown in Table 4, significant
improvements were observed across all metrics when the transfer learning strategy using
the COCO dataset as the source domain was implemented.

Table 4. Impact of transfer learning.

Training Strategy P R F1 mAP50 mAP50-95

Without transfer learning 0.660 0.432 0.52 0.477 0.211
With transfer learning 0.752 0.519 0.60 0.566 0.271

The positive impact of using the COCO dataset as the source domain can be attributed
to several factors. Firstly, the COCO dataset contains a variety of images with diverse and
complex backgrounds, mirroring the conditions often encountered in transmission line
inspection tasks. Secondly, the COCO dataset covers 80 common object categories in daily
life with over 2 million annotated boxes. Many object features and edge characteristics in
the COCO dataset share similarities with features found in transmission line components.
Consequently, the application of transfer learning from the COCO dataset to the
transmission line inspection dataset proves to be an effective strategy, enhancing both the
performance and robustness of the TLI-YOLOv5 model.

3.3.3. Selection of Learning Rate Adjustment Strategy

The performance comparison of models trained with two different learning rate decay
strategies, linear learning rate decay and cosine learning rate decay, is presented in Table 5.
Notably, the model trained with linear learning rate decay exhibits superior precision.
However, the model trained with cosine learning rate decay outperforms the linear learning
rate decay in recall, F1 score, mAP50, and mAP50-95 metrics. Overall, the model trained
with the cosine learning rate decay strategy shows better comprehensive performance.

Table 5. Performance comparison of models trained with linear learning rate decay (linear-lr) and
cosine learning rate decay (cos-lr).

Training Strategy P R F1 mAP50 mAP50-95

With linear-lr 0.760 0.500 0.59 0.551 0.269
With cos-lr 0.752 0.519 0.60 0.566 0.271

To further investigate the impact of cosine learning rate decay on our model, we
incorporated it as an independent component in the ablation experiments of TLI-YOLOv5,
as detailed in Section 3.3.4.
Electronics 2023, 12, 3340 17 of 25

3.3.4. Ablation Experiments of TLI-YOLOv5

To gain insights into the contribution of each component—the SimAM attention
module, the WIoU loss function, and the cosine learning rate decay strategy—we conducted
a total of eight experiments, including different combinations of these components. The
results of these experiments are summarized in Table 6.

Table 6. Ablation results of each component of TLI-YOLOv5.

SimAM WIoU cos-lr P R F1 mAP50 mAP50-95 FPS GFLOPs

√ 0.749 0.499 0.59 0.550 0.269 76.9 4.2

√ 0.748 0.501 0.59 0.554 0.270 76.1 4.2
√ 0.707 0.520 0.59 0.551 0.270 76.6 4.2
√ √ 0.724 0.502 0.58 0.556 0.268 76.9 4.2
√ √ 0.760 0.500 0.59 0.551 0.269 75.2 4.2
√ √ 0.723 0.516 0.59 0.561 0.274 75.2 4.2
√ √ √ 0.766 0.509 0.61 0.560 0.270 76.9 4.2
0.752 0.519 0.60 0.566 0.271 76.1 4.2

The baseline experiment employs the original YOLOv5n model without any of
these enhancements.
The experimental results indicate that when the SimAM attention module and the
WIoU loss function are independently implemented, they both result in marginal improve-
ments to the overall performance. However, the standalone application of the cosine
learning rate decay strategy exhibits a slight decrement in performance. Yet, it should be
noted that these effects are quite subtle.
Upon combining any two enhancements—either SimAM and WIoU, SimAM and
cos-lr, or WIoU and cos-lr—the performance measures notably surpass those when any
single component is applied independently. This suggests that the amalgamation of these
techniques leads to a synergistic boost in model performance.
Remarkably, the simultaneous incorporation of all three modifications yields peak
performance across all combinations, illustrating a potential synergistic effect. Specifically,
improvements are seen as increases of 0.40% in precision, 4.01% in recall, 1.69% in the F1
score, 2.91% in mAP50, and 0.74% in mAP50-95, respectively.
Importantly, the implementation of these three enhancements does not alter the
GFLOPs value of the model, and the model’s weight file retains the same 4.15 MB size
as the original YOLOv5n. The difference in FPS is negligible. This indicates that the TLI-
YOLOv5 framework successfully maintains the lightweight characteristics of YOLOv5n
while simultaneously improving detection performance.

3.3.5. Comparison with Mainstream Networks

A comparative analysis was conducted between our proposed TLI-YOLOv5 and
several mainstream networks using our transmission line inspection dataset.
Specifically, we included YOLOv3 in the comparison since it was used in the stud-
ies [33–35] for transmission line inspection. Additionally, we selected YOLOv3-tiny, a
lightweight version of YOLOv3, aimed at achieving faster inference speed at the cost
of some accuracy. Furthermore, we incorporated the latest lightweight algorithms from
the YOLO series, namely YOLOv6n and YOLOv8n (YOLOv7 currently does not have a
lightweight variant available), into our comparison. YOLOv6n is renowned for its ex-
ceptional performance with respect to both accuracy and efficiency. On the other hand,
YOLOv8n is regarded as the state-of-the-art lightweight variant among the YOLO series.
The results are summarized in Table 7.
Electronics 2023, 12, 3340 18 of 25

Table 7. Comparison with mainstream networks.

Model P R mAP50 mAP50-95 FPS GFLOPs

YOLOv3 0.795 0.566 0.625 0.318 15.8 154.7
YOLOv3-tiny 0.588 0.365 0.385 0.159 72.5 12.9
YOLOv6n 0.704 0.421 0.453 0.249 74.6 11.4
YOLOv8n 0.753 0.432 0.475 0.255 76.9 8.9
TLI-YOLOv5 0.752 0.519 0.566 0.271 76.1 4.2

The results clearly demonstrate the advantages and characteristics of each network
in terms of accuracy, speed, and model size. Notably, our proposed TLI-YOLOv5 model
achieves competitive performance.
Overall, these comparative analyses highlight the effectiveness of TLI-YOLOv5 as a
promising object detection framework for transmission line inspection, offering a balanced
combination of accuracy and efficiency for real-world applications.

4. Discussion
Table 8 displays the detection performance of TLI-YOLOv5 for eight different objects.
The confusion matrix of TLI-YOLOv5’s detection results is illustrated in Figure 11. The
x-axis represents the actual categories, while the y-axis indicates the predicted categories.
Each cell in the matrix represents the proportion of instances where the model predicted a
given actual category as the predicted category. All proportion values in the matrix have
been normalized to the range of 0 to 1.

Table 8. Performance of TLI-YOLOv5 on different object recognition.

Object P R mAP50 mAP50-95

All 0.752 0.519 0.566 0.271
Tower 0.755 0.594 0.699 0.340
Insulator 0.753 0.469 0.542 0.263
Plate 0.668 0.504 0.542 0.252
Ring 0.732 0.283 0.348 0.183
Clamp 0.734 0.534 0.533 0.316
Damper 0.870 0.565 0.629 0.310
Sign 0.680 0.341 0.377 0.175
Nest 0.822 0.864 0.860 0.328

Among all eight categories, the model performs best in detecting nest, followed by
vibration damper and tower. The detection results for insulator, yoke plate, and line clamp
are at an average level. In contrast, the worst-performing categories are corona ring and
tower sign, whose performance is significantly lower than the other six objects.
Delving into the images of these two problematic categories, we can identify the
contributing factors to these discrepancies.
For the corona ring, it holds the position of the smallest object among all eight cat-
egories. Its small size inherently hampers detection effectiveness. Specifically, smaller
objects occupy fewer pixels in an image, making it more challenging for the model to learn
and extract meaningful features.
As for the tower sign, in some images captured at mid to long distances, tower signs
still have high recognizability due to their distinctive features, despite their small size.
Thus, we annotated them in the dataset. However, similar to other small objects, the model
faces difficulty accurately detecting them. As illustrated in Figure 12, among the seven
tower signs that we annotated in the original image, only one was successfully detected by
the model.
Electronics 2023, 12, 3340 19 of 25
Electronics 2023, 12, x FOR PEER REVIEW 20 of 26

Figure 11. Confusion matrix.

Figure 11. Confusion matrix.

Among
Figure 13all eight categories,
provides the model
a clear visual performs
comparison bestperformance
of the in detecting differences
nest, followed by
between
vibration damper and tower. The detection results for insulator, yoke plate,
the original YOLOv5n and TLI-YOLOv5. It is evident that the confidence threshold of the and line clamp
are at an bounding
detection average level. In contrast,
boxes the worst-performing
by TLI-YOLOv5 is generally categories
higher thanarethat
corona ringoriginal
of the and
tower sign, whose performance is significantly lower than the other six
YOLOv5n. TLI-YOLOv5 successfully detects the largest object in the image, the tower,objects.
Delving into the images of these two problematic categories, we can identify the con-
which YOLOv5n fails to detect.
tributing factors to these discrepancies.
Figure 14 presents the detailed sections of Figure 13. As demonstrated in Figure 14a,
For the corona ring, it holds the position of the smallest object among all eight cate-
TLI-YOLOv5 successfully detected the obscured object hidden behind the tower body,
gories. Its small size inherently hampers detection effectiveness. Specifically, smaller ob-
which the original YOLOv5n was unable to detect. Furthermore, as illustrated in Figure 14b,
jects occupy fewer pixels in an image, making it more challenging for the model to learn
TLI-YOLOv5 generates more accurate detection bounding boxes than the original YOLOv5n
and extract meaningful features.
for densely arranged objects.
As for the tower sign, in some images captured at mid to long distances, tower signs
As a lightweight object detection framework designed for real-time detection and
still have high recognizability due to their distinctive features, despite their small size.
video
Thus,detection, TLI-YOLOv5,
we annotated them in thesimilar
dataset.toHowever,
other lightweight models,
similar to other compromises
small a degree
objects, the model
offaces
precision for a higher detection speed. In the context of UAV transmission
difficulty accurately detecting them. As illustrated in Figure 12, among the seven line inspection,
the UAV’s
tower visuals
signs that weconstitute
annotateda continuous
in the originalvideo stream.
image, only The speed
one was of detection
successfully not only
detected
determines the
by the model. fluidity of the UAV’s operation but also significantly affects the efficiency
of the inspection work. Through testing, TLI-YOLOv5’s performance metrics satisfy the
demands of such tasks, striking a well-managed balance between precision and speed.
2, x FOR PEER REVIEW 21 of 26
Electronics 2023, 12, 3340 20 of 25

Figure 12. Detection result

Figure 12. of tower result
Detection signsofintower
a sample
signs inimage: (a)image:
a sample original annotations;
(a) original (b) detection
annotations; (b) detection
results. results.

Figure 13 provides a clear visual comparison of the performance differences between

the original YOLOv5n and TLI-YOLOv5. It is evident that the confidence threshold of the
detection bounding boxes by TLI-YOLOv5 is generally higher than that of the original
YOLOv5n. TLI-YOLOv5 successfully detects the largest object in the image, the tower,
which YOLOv5n fails to detect.
 Electronics
Electronics 2023,
2023, 12,12, x FOR PEER REVIEW
3340 22 of
21 26
of 25

Figure 13. Comparison of detection results between YOLOv5n and TLI-YOLOv5: (a) YOLOv5n; (b)
TLI-YOLOv5.

Figure 14 presents the detailed sections of Figure 13. As demonstrated in Figure 14a,
TLI-YOLOv5 successfully detected the obscured object hidden behind the tower body,
which the original YOLOv5n was unable to detect. Furthermore, as illustrated in Figure
14b,
FigureTLI-YOLOv5
Figure 13. generates
13. Comparison
Comparison of more
ofdetection
detection accurate
results detection
between
results YOLOv5n
between bounding
andand
YOLOv5n boxes (a)
TLI-YOLOv5: than
TLI-YOLOv5: theYOLOv5n;
original
YOLOv5n;
(a) (b)
YOLOv5n
TLI-YOLOv5. for
(b) TLI-YOLOv5. densely arranged objects.

Figure 14 presents the detailed sections of Figure 13. As demonstrated in Figure 14a,
TLI-YOLOv5 successfully detected the obscured object hidden behind the tower body,
which the original YOLOv5n was unable to detect. Furthermore, as illustrated in Figure
14b, TLI-YOLOv5 generates more accurate detection bounding boxes than the original
YOLOv5n for densely arranged objects.

Figure 14.
Figure 14. Comparison of detection
detection results:
results: (a) obscured
obscured object;
object; (b)
(b) densely
densely arranged
arranged objects.
objects.

To further analyze the effectiveness of our proposed TLI-YOLOv5, we compare it

with some other existing object detection methods designed for UAV-based transmission
line inspection.
Figure 14. Comparison
Previous works of detection
[33–35] results: (a)
adopted obscuredas
YOLOv3 object;
the (b) densely arranged
detection objects.
algorithm. However,
YOLOv3 has been widely recognized to be surpassed by YOLOv5 in comprehensive perfor-
mance, which is also validated in our experiments in Section 3.3.5. Previous works [36,37]
employed earlier versions of YOLOv5, whose capabilities fall far behind the 7.0 version
Electronics 2023, 12, 3340 22 of 25

used in our study. Moreover, they did not utilize the most lightweight YOLOv5n model,
thus their detection speed is much lower compared to our approach.
Most importantly, our dataset contains eight types of transmission line objects and is
annotated based on real-world UAV inspection images. Therefore, our proposed method
has a wider range of application scenarios and is more suitable for the needs of modern un-
manned aerial vehicle transmission line inspection. It demonstrates noticeable superiority
over existing approaches.
Despite its strengths, our work has certain limitations. First, transmission line inspec-
tions involve a wide array of components. Due to constraints of time and manpower, our
research only annotated and trained eight types of components. In future work, we plan to
incorporate more components into our detection targets. Additionally, we aim to augment
our dataset with a diverse range of medium- and large-sized objects, enabling the model
to learn and capture more intricate feature details. Second, with the continuous introduc-
tion of more sophisticated algorithms and improvement strategies in the field of object
detection, it is worth exploring their application to UAV transmission line inspection tasks.
For example, recent works such as [47,48] on network pruning could help reduce model
complexity. Techniques such as weight quantization [49,50] may be adopted for significant
memory reduction. We hope our work can provide a baseline for future explorations to
further advance UAV-based transmission line inspection capabilities.

5. Conclusions
This study proposes TLI-YOLOv5, a lightweight object detection framework specially
developed for transmission line inspection. By integrating the parameter-free attention
module SimAM, the WIoU loss function, and incorporating advanced training strategies,
TLI-YOLOv5 demonstrates enhanced performance in various metrics compared to the
original YOLOv5n, while maintaining high recognition speed and compactness. This work
contributes a practical and efficient solution for large-scale, real-time transmission line
inspection tasks.
However, there remain some limitations in our current research. First, the number
of object categories is still limited. In future work, more types of components could
be incorporated as detection targets to expand the model’s capabilities. Second, While
TLI-YOLOv5 achieves a balance between accuracy and efficiency, using more advanced
networks could potentially further boost detection performance. We aim to explore state-
of-the-art architectures to continue pushing the frontiers. Third, creating larger datasets
covering more environments would also be beneficial. Overall, this study offers a robust
baseline model, and future research will focus on expanding in terms of category diversity,
maximizing accuracy, and improving generalizability.
Although object detection has made rapid progress, there remain challenges in inte-
gration and implementation within real-world transmission line inspection applications.
Factors such as dataset limitations, model generalization, efficiency constraints, and transi-
tion costs can hinder adoption. Our future work will concentrate on tackling these pitfalls
to facilitate the broader deployment of object detection algorithms in transmission line
inspection systems. We believe continued research and technical development efforts will
progressively bridge the gap between state-of-the-art techniques and practical usage in the
field of transmission line inspection.

Author Contributions: Conceptualization, H.H. and G.L.; methodology, H.H. and J.W.; software,
H.H. and Z.Z.; validation, H.H., G.L., J.W., Z.Z., Z.X., D.L. and F.Z.; formal analysis, H.H., G.L.
and J.W.; investigation, H.H., J.W. and D.L.; resources, G.L.; data curation H.H., Z.Z., Z.X. and F.Z.;
writing—original draft preparation, H.H.; writing—review and editing, H.H. and G.L.; visualization,
Z.Z., Z.X. and D.L.; supervision, G.L.; project administration, G.L.; funding acquisition, G.L. All
authors have read and agreed to the published version of the manuscript.
Funding: This work was supported by a grant from the National Natural Science Foundation of
China (Grant number 41861050).
Electronics 2023, 12, 3340 23 of 25

Data Availability Statement: The data are not publicly available due to the confidentiality of the
research projects.
Acknowledgments: The authors extend their profound gratitude to the anonymous reviewers for
their insightful critiques and constructive suggestions on the manuscript.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Luo, Y.; Yu, X.; Yang, D.; Zhou, B. A survey of intelligent transmission line inspection based on unmanned aerial vehicle. Artif.
Intell. Rev. 2023, 56, 173–201. [CrossRef]
2. Alhassan, A.B.; Zhang, X.; Shen, H.; Xu, H. Power transmission line inspection robots: A review, trends and challenges for future
research. Int. J. Electr. Power Energy Syst. 2020, 118, 105862. [CrossRef]
3. Larrauri, J.I.; Sorrosal, G.; González, M. Automatic system for overhead power line inspection using an Unmanned Aerial
Vehicle—RELIFO project. In Proceedings of the 2013 International Conference on Unmanned Aircraft Systems (ICUAS), Atlanta,
GA, USA, 28–31 May 2013; pp. 244–252.
4. Luque-Vega, L.F.; Castillo-Toledo, B.; Loukianov, A.; Gonzalez-Jimenez, L.E. Power line inspection via an unmanned aerial system
based on the quadrotor helicopter. In Proceedings of the MELECON 2014—2014 17th IEEE Mediterranean Electrotechnical
Conference, Beirut, Lebanon, 13–16 April 2014; pp. 393–397.
5. Matikainen, L.; Lehtomäki, M.; Ahokas, E.; Hyyppä, J.; Karjalainen, M.; Jaakkola, A.; Kukko, A.; Heinonen, T. Remote sensing
methods for power line corridor surveys. ISPRS J. Photogramm. Remote Sens. 2016, 119, 10–31. [CrossRef]
6. Wang, C.-N.; Yang, F.-C.; Vo, N.T.; Nguyen, V.T.T. Wireless communications for data security: Efficiency assessment of cybersecu-
rity industry—A promising application for UAVs. Drones 2022, 6, 363. [CrossRef]
7. Wang, C.-N.; Yang, F.-C.; Vo, N.T.; Nguyen, V.T.T. Enhancing Lithium-Ion Battery Manufacturing Efficiency: A Comparative
Analysis Using DEA Malmquist and Epsilon-Based Measures. Batteries 2023, 9, 317. [CrossRef]
8. Bian, J.; Hui, X.; Zhao, X.; Tan, M. A novel monocular-based navigation approach for UAV autonomous transmission-line
inspection. In Proceedings of the 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Madrid,
Spain, 1–5 October 2018; pp. 1–7.
9. Li, X.; Li, Z.; Wang, H.; Li, W. Unmanned aerial vehicle for transmission line inspection: Status, standardization, and perspectives.
Front. Energy Res. 2021, 9, 713634. [CrossRef]
10. Hui, X.; Bian, J.; Zhao, X.; Tan, M. Vision-based autonomous navigation approach for unmanned aerial vehicle transmission-line
inspection. Int. J. Adv. Robot. Syst. 2018, 15, 1729881417752821. [CrossRef]
11. Mirallès, F.; Pouliot, N.; Montambault, S. State-of-the-art review of computer vision for the management of power transmission
lines. In Proceedings of the 2014 3rd International Conference on Applied Robotics for the Power Industry, Foz do Iguacu, Brazil,
14–16 October 2014; pp. 1–6.
12. Jenssen, R.; Roverso, D. Automatic autonomous vision-based power line inspection: A review of current status and the potential
role of deep learning. Int. J. Electr. Power Energy Syst. 2018, 99, 107–120.
13. Han, Y.; Liu, Z.; Lee, D.; Liu, W.; Chen, J.; Han, Z. Computer vision–based automatic rod-insulator defect detection in high-speed
railway catenary system. Int. J. Adv. Robot. Syst. 2018, 15, 1729881418773943. [CrossRef]
14. Ma, Y.; Li, Q.; Chu, L.; Zhou, Y.; Xu, C. Real-time detection and spatial localization of insulators for UAV inspection based on
binocular stereo vision. Remote Sens. 2021, 13, 230. [CrossRef]
15. Girshick, R.; Donahue, J.; Darrell, T.; Malik, J. Rich feature hierarchies for accurate object detection and semantic segmentation.
In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Columbus, OH, USA, 23–28 June 2014;
pp. 580–587.
16. Ren, S.; He, K.; Girshick, R.; Sun, J. Faster r-cnn: Towards real-time object detection with region proposal networks. Adv. Neural
Inf. Process. Syst. 2015, 28, 91–99. [CrossRef]
17. He, K.; Gkioxari, G.; Dollár, P.; Girshick, R. Mask r-cnn. In Proceedings of the IEEE International Conference on Computer Vision,
Venice, Italy, 22–29 October 2017; pp. 2961–2969.
18. Redmon, J.; Divvala, S.; Girshick, R.; Farhadi, A. You only look once: Unified, real-time object detection. In Proceedings of the
IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas, NV, USA, 27–30 June 2016; pp. 779–788.
19. Liu, W.; Anguelov, D.; Erhan, D.; Szegedy, C.; Reed, S.; Fu, C.-Y.; Berg, A.C. Ssd: Single shot multibox detector. In Proceedings of
the Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, 11–14 October 2016; Proceedings,
Part I 14, 2016. pp. 21–37.
20. Dewi, C.; Chen, R.-C.; Liu, Y.-T.; Jiang, X.; Hartomo, K.D. Yolo V4 for advanced traffic sign recognition with synthetic training
data generated by various GAN. IEEE Access 2021, 9, 97228–97242. [CrossRef]
21. Lan, W.; Dang, J.; Wang, Y.; Wang, S. Pedestrian detection based on YOLO network model. In Proceedings of the 2018 IEEE
International Conference on Mechatronics and Automation (ICMA), Changchun, China, 5–8 August 2018; pp. 1547–1551.
22. Laroca, R.; Severo, E.; Zanlorensi, L.A.; Oliveira, L.S.; Gonçalves, G.R.; Schwartz, W.R.; Menotti, D. A robust real-time automatic
license plate recognition based on the YOLO detector. In Proceedings of the 2018 International Joint Conference on Neural
Networks (IJCNN), Shah Alam, Malaysia, 7 August 2018; pp. 1–10.
Electronics 2023, 12, 3340 24 of 25

23. Liu, G.; Nouaze, J.C.; Touko Mbouembe, P.L.; Kim, J.H. YOLO-tomato: A robust algorithm for tomato detection based on YOLOv3.
Sensors 2020, 20, 2145. [CrossRef] [PubMed]
24. Liu, J.; Wang, X. Tomato diseases and pests detection based on improved Yolo V3 convolutional neural network. Front. Plant Sci.
2020, 11, 898. [CrossRef] [PubMed]
25. Wu, D.; Lv, S.; Jiang, M.; Song, H. Using channel pruning-based YOLO v4 deep learning algorithm for the real-time and accurate
detection of apple flowers in natural environments. Comput. Electron. Agric. 2020, 178, 105742. [CrossRef]
26. Tian, Y.; Yang, G.; Wang, Z.; Wang, H.; Li, E.; Liang, Z. Apple detection during different growth stages in orchards using the
improved YOLO-V3 model. Comput. Electron. Agric. 2019, 157, 417–426. [CrossRef]
27. Loey, M.; Manogaran, G.; Taha, M.H.N.; Khalifa, N.E.M. Fighting against COVID-19: A novel deep learning model based on
YOLO-v2 with ResNet-50 for medical face mask detection. Sustain. Cities Soc. 2021, 65, 102600. [CrossRef]
28. Punn, N.S.; Sonbhadra, S.K.; Agarwal, S.; Rai, G. Monitoring COVID-19 social distancing with person detection and tracking via
fine-tuned YOLO v3 and Deepsort techniques. arXiv 2020, arXiv:2005.01385.
29. Cheng, L.; Li, J.; Duan, P.; Wang, M. A small attentional YOLO model for landslide detection from satellite remote sensing im
ages. Landslides 2021, 18, 2751–2765. [CrossRef]
30. Du, Y.; Pan, N.; Xu, Z.; Deng, F.; Shen, Y.; Kang, H. Pavement distress detection and classification based on YOLO network. Int. J.
Pavement Eng. 2021, 22, 1659–1672. [CrossRef]
31. Aly, G.H.; Marey, M.; El-Sayed, S.A.; Tolba, M.F. YOLO based breast masses detection and classification in full-field digital
mammograms. Comput. Methods Programs Biomed. 2021, 200, 105823. [CrossRef] [PubMed]
32. Ünver, H.M.; Ayan, E. Skin lesion segmentation in dermoscopic images with combination of YOLO and grabcut algorithm.
Diagnostics 2019, 9, 72. [CrossRef] [PubMed]
33. Chen, H.; He, Z.; Shi, B.; Zhong, T. Research on recognition method of electrical components based on YOLO V3. IEEE Access
2019, 7, 157818–157829. [CrossRef]
34. Chen, B.; Miao, X. Distribution line pole detection and counting based on YOLO using UAV inspection line video. J. Electr. Eng.
Technol. 2020, 15, 441–448. [CrossRef]
35. Renwei, T.; Zhongjie, Z.; Yongqiang, B.; Ming, G.; Zhifeng, G. Key parts of transmission line detection using improved YOLO v3.
Int. Arab J. Inf. Technol. 2021, 18, 747–754. [CrossRef]
36. Zhan, W. Electric Equipment Inspection on High Voltage Transmission Line Via Mobile Net-SSD. CONVERTER 2021, 2021,
527–540.
37. Bao, W.; Du, X.; Wang, N.; Yuan, M.; Yang, X. A Defect Detection Method Based on BC-YOLO for Transmission Line Components
in UAV Remote Sensing Images. Remote Sens. 2022, 14, 5176. [CrossRef]
38. Jocher, G.; Chaurasia, A.; Stoken, A.; Borovec, J.; Kwon, Y.; Michael, K.; Fang, J.; Yifu, Z.; Wong, C.; Montes, D. Ultralytics/yolov5:
v7. 0-yolov5 Sota Realtime Instance Segmentation. Zenodo. 2022. Available online: https://zenodo.org/record/7347926
(accessed on 4 July 2023).
39. Jocher, G.; Stoken, A.; Chaurasia, A.; Borovec, J.; Kwon, Y.; Michael, K.; Changyu, L.; Fang, J.; Skalski, P.; Hogan, A. ultralyt-
ics/yolov5: v6. 0-Yolov5n’nano’models, Roboflow Integration, TensorFlow Export, OpenCV DNN support. Zenodo. 2021.
Available online: https://zenodo.org/record/5563715 (accessed on 4 July 2023).
40. Yang, L.; Zhang, R.-Y.; Li, L.; Xie, X. Simam: A simple, parameter-free attention module for convolutional neural networks. In
Proceedings of the International Conference on Machine Learning, Virtual, 18–24 July 2021; pp. 11863–11874.
41. Tong, Z.; Chen, Y.; Xu, Z.; Yu, R. Wise-IoU: Bounding Box Regression Loss with Dynamic Focusing Mechanism. arXiv 2023,
arXiv:2301.10051.
42. Zhuang, F.; Qi, Z.; Duan, K.; Xi, D.; Zhu, Y.; Zhu, H.; Xiong, H.; He, Q. A comprehensive survey on transfer learning. Proc. IEEE
2020, 109, 43–76. [CrossRef]
43. Zheng, J.; Lu, C.; Hao, C.; Chen, D.; Guo, D. Improving the generalization ability of deep neural networks for cross-domain visual
recognition. IEEE Trans. Cogn. Dev. Syst. 2020, 13, 607–620. [CrossRef]
44. Hao, C.; Chen, D. Software/hardware co-design for multi-modal multi-task learning in autonomous systems. In Proceedings of
the 2021 IEEE 3rd International Conference on Artificial Intelligence Circuits and Systems (AICAS), Washington DC, USA, 6–9
June 2021; pp. 1–5.
45. He, T.; Zhang, Z.; Zhang, H.; Zhang, Z.; Xie, J.; Li, M. Bag of tricks for image classification with convolutional neural networks. In
Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, Long Beach, CA, USA, 15–20 June 2019;
pp. 558–567.
46. Zheng, Z.; Wang, P.; Liu, W.; Li, J.; Ye, R.; Ren, D. Distance-IoU loss: Faster and better learning for bounding box regression. In
Proceedings of the AAAI Conference on Artificial Intelligence, Virtual, 22 February–1 March 2020; pp. 12993–13000.
47. Hu, W.; Che, Z.; Liu, N.; Li, M.; Tang, J.; Zhang, C.; Wang, J. Channel Pruning via Class-Aware Trace Ratio Optimization. IEEE
Trans. Neural Netw. Learn. Syst. 2023. [CrossRef] [PubMed]
48. Tang, Z.; Luo, L.; Xie, B.; Zhu, Y.; Zhao, R.; Bi, L.; Lu, C. Automatic sparse connectivity learning for neural networks. IEEE Trans.
Neural Netw. Learn. Syst. 2022. [CrossRef] [PubMed]
Electronics 2023, 12, 3340 25 of 25

49. Huang, Q. Weight-quantized squeezenet for resource-constrained robot vacuums for indoor obstacle classification. AI 2022, 3,
180–193. [CrossRef]
50. Ding, C.; Wang, S.; Liu, N.; Xu, K.; Wang, Y.; Liang, Y. REQ-YOLO: A resource-aware, efficient quantization framework for object
detection on FPGAs. In Proceedings of the 2019 ACM/SIGDA International Symposium on Field-Programmable Gate Arrays,
Seaside, CA, USA, 24–26 February 2019; pp. 33–42.

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.